Hierarchical Framework for Remote Sensing Deep Learning Evaluation on Terminal Devices

The fusion of multi-source remote sensing data (optical, SAR, and infrared) on terminal devices with limited resources poses strict demands on performance evaluation. Existing approaches provide no unified or systematic framework for assessing deep learning models under heterogeneous data modalities and hardware constraints. We propose a four-layer hierarchical evaluation framework consisting of the mode, device, model, and dataset levels. Within this framework, the Hybrid Hierarchical Evaluation Algorithm (HHEA) employs a multi-stage fusion scheme across layers to enable rigorous analysis across different dimensions. The framework supports benchmark tests on representative deep learning workloads and evaluates lightweight models under both single-thread and multi-thread execution. Extensive experiments on diverse terminals, models, and datasets demonstrate that HHEA effectively identifies performance bottlenecks and informs task scheduling and model deployment in terminal-oriented remote sensing applications. Moreover, the framework is readily extensible to additional tasks, providing deeper insights into the adaptability between models and hardware architectures.